Where to go next
Review the core vocabulary, build a complete tiny classifier end to end, and pick your next track.
- Use the core AI vocabulary with confidence
- Combine training, prediction and evaluation in one program
- Choose a next step that fits your goals
You’ve gone from “what is AI?” to vectors, models, neural networks, search and games, rewards and generative AI, vision, language and ethics. Here’s the vocabulary you’ve picked up - keep it handy.
| Term | Meaning |
|---|---|
| Feature | A measurable input a model uses to make a prediction. |
| Label | The correct answer attached to a training example. |
| Vector | An ordered list of numbers representing one thing. |
| Dot product | Multiply matching entries of two vectors and add them up. |
| Regression | Predicting a number. |
| Classification | Predicting a category. |
| k-nearest neighbors | Predict by a majority vote of the k closest training examples. |
| Loss function | One number measuring how wrong the predictions are. |
| Overfitting | Memorizing the training data instead of learning a general pattern. |
| Gradient descent | Repeatedly nudging parameters downhill on the loss. |
| Learning rate | The size of each gradient descent step. |
| Epoch | One full pass through the training data. |
| Neural network | Layers of connected artificial neurons. |
| Activation function | The nonlinear function applied to a neuron’s weighted sum. |
| Bias (parameter) | The learned constant added to a weighted sum. |
| Backpropagation | Computing every weight’s gradient by working backward from the output. |
| Deep learning | Machine learning with many-layered neural networks. |
| Convolution | Sliding a small kernel over a grid, taking a weighted sum at each spot. |
| Token | A unit of text (often a subword) that a language model processes. |
| Embedding | A learned vector that places similar items near each other. |
| Hallucination | A fluent, confident, false statement from a generative model. |
| Reinforcement learning | Learning by trial and error from rewards. |
| A* search | Finds the cheapest route by exploring by cost so far plus a guess of the cost to go. |
| Minimax | Picks the move with the best outcome you can guarantee against a perfect opponent. |
| Bayes’ rule | Updates a prior belief into a posterior using new evidence. |
| Q-value | An agent’s estimate of the future reward from taking an action in a state. |
| Diffusion model | A generator that creates images by learning to remove noise step by step. |
| Precision / recall | Of the alarms raised, how many were right / of the real cases, how many were caught. |
Try it
Up to 15 XPQuick vocabulary sort
Which part of building a model does each term belong to?
“Features and labels”
“Activation function”
“Learning rate”
“Held-out test set”
“Embeddings”
“Epochs and batches”
“Per-group error rates”
“Train/test split”
Everything together
Here is a complete, tiny machine learning pipeline: split the data, train (compute each class’s average point, its centroid), predict (nearest centroid), and evaluate on held-out data. It’s a close cousin of both k-NN and k-means. The exercise asks you to write it yourself.
1import math
2
3data = [((1.0, 1.2), "small"), ((1.4, 0.9), "small"), ((0.8, 1.1), "small"), ((1.2, 1.5), "small"),
4 ((4.1, 3.8), "large"), ((3.7, 4.2), "large"), ((4.4, 4.0), "large"), ((3.2, 2.6), "large")]
5train, test = data[:3] + data[4:7], [data[3], data[7]]
6
7centroids = {}
8for label in ("small", "large"):
9 points = [point for point, point_label in train if point_label == label]
10 centroids[label] = tuple(sum(values) / len(values) for values in zip(*points))
11
12def predict(point):
13 return min(centroids, key=lambda label: math.dist(point, centroids[label]))
14
15correct = sum(predict(point) == label for point, label in test)
16print(f"test accuracy: {correct}/{len(test)}")test accuracy: 2/2
Keep learning
Pick the track that matches what excited you most:
- Machine Learning - logistic regression, decision trees, cross-validation, and a full capstone classifier with numpy.
- Data Science - explore, clean and visualize real data with pandas.
- Computer Vision - filters, segmentation, detection and CNNs in depth.
- Natural Language Processing - from tokenizers to word vectors and transformers.
- Large Language Models, Prompt Engineering and RAG - build with the models behind modern chatbots.
And some excellent free resources outside this site:
- Elements of AI - a friendly, no-code introduction from the University of Helsinki.
- Google’s Machine Learning Crash Course - the fundamentals in more depth, with interactive exercises.
- Harvard CS50’s Introduction to AI with Python - a project-based university course.
- 3Blue1Brown’s neural networks series - beautiful visual intuition for networks and backpropagation.
- fast.ai - a practical, code-first deep learning course.
Key takeaways
Every ML project follows the same shape: data → model → training → evaluation.
The core ideas - vectors, dot products, loss, gradient descent - reappear in every corner of AI.
Keep going with a deeper track, and build something small to make it stick.
Lesson quiz
6 questions · pass with 5 correct · up to 50 XP
Passing this quiz completes the lesson and keeps your streak going. Questions you miss come back in review sessions later.
Practice: write Python
Write Python in the editor and run it against sample inputs. Python runs locally in your browser using a WebAssembly runtime.
Build a complete classifier
Each input line is x y label. The last 25% of lines (rounding down the number of training lines to int(count * 0.75)) are the test set; the rest are for training.
- Train: for each label, compute the centroid (average x, average y) of its training points. Print them in alphabetical order as
LABEL: (X, Y)with 2 decimal places. - Predict each test point’s label as the label of the nearest centroid, and print
x y -> PREDICTED (actual ACTUAL)(format x and y with:g). - Print
accuracy: P%(whole percent).
When it works, look at the mistake it makes on the fruit test: weight in grams (100-170) swamps the color score (2-8) in the distance - the scaling pitfall from the k-NN lesson.
- Fruit
Python runs in a sandboxed browser worker with a 60 second time limit. Its runtime loads from the Pyodide CDN; your code stays in this browser.
Questions about this lesson
Stuck? Ask. Figured something out? Share it. Explaining is one of the best ways to learn.
Loading posts…